Medical image feature selection method based on population tree search and hybrid intelligent heredity

Abstract

The invention discloses a medical image feature selection method based on population tree search and hybrid intelligent heredity. The method comprises the following steps: acquiring a to-be-processed medical image feature data set; initializing a population tree; evaluating and calculating the fitness; and executing branch evolution, starting from the current tree node, and generating a plurality of sub-generation branches in parallel. Each branch generates a new population by applying a hybrid genetic algorithm to the parent medical image feature subset population; analyzing statistical summary information of all filial generation branches by using a decision model; and performing depth-first search on the population tree until a preset termination condition is met. And finally, outputting a medical image feature subset with the highest fitness found in the whole search process as an optimal feature selection result. According to the method, the problems that a traditional feature selection method is prone to falling into local optimum in high-dimensional data and lacks intelligent guidance are solved, and the global search capability, the convergence speed and the solution quality of the feature selection process are improved.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided diagnostic technology, specifically relating to a feature selection method and system for medical image analysis. More particularly, it relates to a method that utilizes large language models and population tree search to select the optimal feature subset from high-dimensional medical image quantification features, thereby improving the performance of downstream diagnostic models. Background Technology

[0002] In today's era of big data, the rapid growth of data brings new challenges. On the one hand, with the digital transformation of fields such as healthcare and industrial manufacturing, data acquisition capabilities have been unprecedentedly enhanced, leading to a dramatic expansion of data dimensions and generating a large amount of noise, redundancy, and even erroneous information irrelevant to the learning task, resulting in the "curse of dimensionality." On the other hand, enterprises and research institutions have increasingly stringent requirements for the efficiency of model training and deployment. Processing high-dimensional data not only consumes enormous computing and storage resources, but also significantly prolongs the model development cycle, thus limiting the practical application value of algorithms. The existence of these problems necessitates that we must balance model performance and efficiency when dealing with high-dimensional data. This challenge is particularly prominent in the field of medical image analysis. For example, a single CT or SPECT scan image can generate thousands or even tens of thousands of quantitative features such as texture, morphology, and intensity. However, only a small fraction of these massive features are decisive for the diagnosis of specific diseases (such as determining the benign or malignant nature of tumors or diagnosing heart disease). The remaining large number of features not only contain redundant information and noise (such as imaging artifacts and individual differences), but also greatly increase the computational burden of model training and may mislead the model, leading to overfitting and decreased generalization ability. Therefore, how to automatically and efficiently select a subset of key features from high-dimensional medical image features that has both high predictive performance and clinical interpretability is the core challenge in improving the performance of computer-aided diagnostic systems.

[0003] Evolutionary algorithms (EAs) are a class of metaheuristic optimization algorithms inspired by natural selection and genetic mechanisms in the biological world. Their core idea is to simulate the evolutionary process of "survival of the fittest," encompassing three typical methods: genetic algorithms, evolutionary strategies, and evolutionary programming. Genetic algorithms were created by J. Holland in the United States; evolutionary programming was proposed by L.J. Fogel, A.J. Johns, and M.J. Walsh in the United States, and refined by D.B. Fogel; and evolutionary strategies were proposed by I. Reehenberg and H.P. Schwefel in Germany.

[0004] Evolutionary algorithms include the following core definitions:

[0005] An individual refers to a potential solution in the search space of an optimization problem. In feature selection tasks, each individual represents a subset of candidate features.

[0006] Population: A collection of multiple individuals. Algorithms maintain a population to enable parallel exploration of multiple different regions in the search space.

[0007] Fitness function: An evaluation function used to quantitatively assess the merits of each individual. In feature selection, it typically considers two objectives: the predictive performance of the candidate feature subset (such as classification accuracy) and the simplicity of the subset itself (i.e., the number of features), to ensure that the selected features are as concise as possible while maintaining high classification ability.

[0008] Selection: Simulates the "survival of the fittest" principle in nature. This operator selects superior individuals from the current population as parents based on their fitness values, either probabilistically or deterministically, to generate the next generation of the population.

[0009] Crossover: This operator simulates the gene recombination process in biological heredity. It selects two or more individuals from the parent population and generates one or more new offspring individuals by exchanging parts of their codes (i.e., partial features in feature selection).

[0010] Mutation: This operator simulates gene mutation in biological heredity. It randomly perturbs the encoding of an individual with a small probability, such as randomly adding or deleting a feature during feature selection. Its main function is to maintain population diversity and prevent the algorithm from prematurely converging to a local optimum.

[0011] Evolutionary algorithms and feature selection tasks are highly complementary. Feature selection requires finding the optimal combination of features within a massive search space, which cannot be solved by brute-force enumeration. The advantage of evolutionary algorithms lies in their global search capability achieved by maintaining a diverse population, allowing them to explore different regions of the search space and effectively reduce the risk of getting trapped in local optima. Furthermore, their black-box optimization nature allows for the flexible design of complex evaluation criteria, incorporating multiple evaluation dimensions such as model accuracy and feature count. Most importantly, by directly evaluating the performance of the entire feature subset, evolutionary algorithms can uncover complex nonlinear dependencies between features.

[0012] However, traditional evolutionary algorithms also have inherent drawbacks that cannot be ignored. As a wrapper method, its computational cost is extremely high because evaluating each individual typically requires training and validating a complete machine learning model, which is particularly time-consuming when dealing with large-scale data. More importantly, its core exploration process relies heavily on random crossover and mutation operations, exhibiting a blind search. The algorithm itself lacks the ability to learn from historical iterations or utilize problem knowledge to guide the search direction. This not only leads to low evolutionary efficiency and a tendency to miss high-quality solutions, but also causes it to get trapped in local optima due to premature loss of population diversity, making it difficult to escape.

[0013] In recent years, artificial intelligence technologies, represented by Large Language Models (LLMs), have made groundbreaking progress, providing new opportunities for the intelligent transformation of evolutionary algorithms. LLMs, leveraging their world knowledge gained through pre-training on massive amounts of data and their powerful contextual understanding and logical reasoning capabilities, can make comprehensive judgments based on task context and historical information to generate high-quality answers. This precisely compensates for the shortcomings of traditional evolutionary algorithms, which rely on random operations and cannot utilize historical knowledge to guide population generation. By embedding LLMs into evolutionary algorithms, genetic operations such as selection, crossover, and mutation can become more targeted, thereby accelerating convergence and improving solution quality. However, most existing research simply applies LLMs to linear, single-path iterative frameworks. This approach has inherent flaws: once the early evolutionary direction deviates, population diversity rapidly decreases, leading to premature convergence to local optima; simultaneously, it lacks the ability to explore in multiple directions in parallel, failing to fully compare the potential of different evolutionary paths.

[0014] Tree of Thoughts (ToT) is an emerging reasoning framework proposed in 2023 by Shunyu Yao et al. from Princeton University. ToT transforms the linear reasoning process of large models into a tree structure, allowing the model to explore, evaluate, and backtrack multiple reasoning paths in parallel, aiming to significantly improve its ability to solve complex problems. Therefore, we integrate ToT into an LLM-driven evolutionary algorithm. ToT expands the originally singular evolutionary process into a multi-branch search tree, where each branch represents an independent, parallel evolutionary direction. This structured multi-path exploration mechanism not only empowers the algorithm with the ability to compare and select between different evolutionary paths through parallel exploration but also introduces a dynamic backtracking mechanism: when the exploration of a branch reaches a bottleneck, the algorithm can backtrack to a parent node with higher potential to re-explore, thereby significantly reducing the risk of premature convergence and greatly enhancing global search capabilities and solution diversity. Summary of the Invention

[0015] The purpose of this invention is to provide an intelligent feature selection method, system, and device for medical image analysis, aiming to solve the problems mentioned in the background art, such as high risk of overfitting, high computational cost, and easy getting trapped in local optima when processing high-dimensional medical image features.

[0016] To achieve the above objectives, the present invention provides the following technical solution:

[0017] A medical image feature selection method based on population tree search and hybrid intelligent genetics includes the following steps:

[0018] S1: Obtain the medical image feature dataset to be processed. The medical image feature dataset to be processed contains multiple samples, and each sample consists of a set of medical image quantitative features and corresponding diagnostic labels.

[0019] S2: Initialize a population tree, where the root node contains multiple randomly generated initial medical image feature subsets, each feature subset serving as an individual in the population;

[0020] S3: Evaluate and calculate fitness. Using a pre-defined fitness function, evaluate the fitness of each subset of medical image features in the population. The fitness function comprehensively considers the predictive performance (e.g., AUC score) of the diagnostic model trained on this feature subset, as well as a penalty term related to the number of selected medical image features, to balance the accuracy and simplicity of the diagnostic model.

[0021] S4: Perform branch evolution, starting from the current tree node, generating multiple child branches in parallel. Each branch generates a new population by applying a hybrid genetic algorithm to the parent medical image feature subset population; the evolutionary operations of the hybrid genetic algorithm include:

[0022] a. Heuristic parent selection: Using a large language model (LLM), parent selection is performed based on indicators such as the fitness of candidate medical image feature subsets;

[0023] b. Intelligent crossover: Prioritize the use of the large language model to perform crossover operations on the parent generation medical image feature subset to generate a child generation medical image feature subset that combines the excellent characteristics of the parents; and fall back to the traditional crossover algorithm under preset conditions (e.g., LLM call failure or meeting a preset probability).

[0024] c. Intelligent mutation: Prioritize the use of the large language model to perform mutation operations on the generated subset of medical image features to explore new feature combinations; and revert to the traditional mutation algorithm under preset conditions;

[0025] S5: Perform intelligent decision-making and tree pruning. Utilize a decision model (e.g., a large language model decision maker) to analyze the statistical summary information of all child branches and select the most promising branch for the next round of evolution; or, if none of the branches have reached the preset potential threshold, perform a backtracking operation to return to the previous node.

[0026] S6: Iteration and Termination. Repeat steps S4 and S5, performing a depth-first search on the population tree until a preset termination condition is met (such as reaching the maximum search depth). The final output is the subset of medical image features with the highest fitness found throughout the search process, which serves as the optimal feature selection result.

[0027] Compared with existing technologies, the present invention has the following innovative features:

[0028] 1. Mind Chain Guided Population Tree Search Strategy: This invention extends the traditional linear evolutionary model into a structured multi-branch population search strategy. Each branch represents an independent parallel evolutionary direction. LLM not only guides the genetic operations within each branch but also performs global evaluation and scheduling between branches through an intelligent decision maker (LLM Voter), realizing path selection, pruning, and backtracking.

[0029] 2. Fitness Function with Dynamic Penalty Mechanism: A dynamic penalty mechanism is introduced into the fitness function that evaluates the quality of feature subsets. When the search process fails to find a better solution for several consecutive generations, the algorithm enters a "stagnation mode" and increases the weight of the feature quantity penalty term, effectively encouraging the algorithm to explore simpler feature combinations and avoid getting trapped in local optima.

[0030] 3. Parallel Evolution Strategy: To address the high time cost associated with multi-branch tree search, this invention introduces a two-layer parallel evolution strategy. At the population level, multiple branches are evolved in parallel to accelerate scheme generation; at the individual level, a feature number adaptive mechanism is employed to perform "segmentation-parallel processing-recombination" operations on high-dimensional feature individuals, thereby significantly improving algorithm efficiency and resource utilization while ensuring optimization quality.

[0031] Compared with existing technologies, the present invention has the following significant advantages:

[0032] 1. Significantly improves global search capability and solution diversity: Through structured population tree search and dynamic backtracking mechanism, the algorithm can maintain the breadth of global search while dynamically concentrating resources on more promising directions and backtracking from invalid paths, which significantly reduces the risk of premature convergence and enhances global search capability and solution diversity.

[0033] 2. Improved convergence speed and efficiency: LLM-driven intelligent operators reduce blind searches and make genetic operations more targeted, thereby accelerating convergence. Simultaneously, the parallel evolution strategy allows populations of different branches to advance simultaneously, improving the utilization of computational resources and overall search efficiency.

[0034] 3. Optimize the balance between model accuracy and simplicity: The balanced fitness function combines the prediction performance of the machine learning model (such as AUC score) with the penalty term for the number of features, and is supplemented by a dynamic penalty mechanism to ensure that the algorithm can generate a compact and accurate subset of features, achieving the best balance between model performance and complexity.

[0035] 4. Enhanced robustness and adaptability of the algorithm: The hybrid genetic algorithm can fall back to the traditional algorithm when the LLM call fails, ensuring the stability of the process. Furthermore, LLM can utilize historical best-case information to assist decision-making, making the genetic operations more informative and further improving the algorithm's adaptability.

[0036] 5. Achieving intelligent and automated feature selection process: This project explores an evolutionary feature selection method driven by a large language model, which is expected to achieve automated evolutionary feature selection and solve the problem of insufficient flexibility and adaptability caused by the need for manual design of evolutionary operation operators in traditional methods. Attached Figure Description

[0037] Figure 1 This is the overall algorithm flowchart.

[0038] Figure 2 This is a flowchart of the population generation process.

[0039] Figure 3 Flowchart for individual processing. Detailed Implementation

[0040] The feature selection system based on population tree search and hybrid intelligent genetic algorithm provided by this invention has the following core components and functions:

[0041] Data loading and preprocessing module: responsible for loading data, handling missing values, data partitioning and standardization.

[0042] Hybrid genetic algorithm engine: performs short-range evolution to produce offspring populations, and internally includes:

[0043] Fitness evaluation unit: Calculates individual fitness, with the main indicator being cross-validation AUC, and applies feature quantity penalty.

[0044] Large Model Crossover / Mutation Unit: Calls LLM to perform intelligent crossover and mutation operations, and supports block processing of high-dimensional data.

[0045] Traditional genetic operator unit: As a backup strategy, it performs traditional crossover and mutation operations when LLM is unavailable or when mixing by strategy.

[0046] Historical solution buffer: Stores historical optimal solutions, providing decision context for LLM operators.

[0047] Population Tree Searcher: Responsible for organizing the evolutionary process into a tree-like search, implementing depth-first advancement and backtracking, and internally includes:

[0048] Branching evolution unit: Multiple sub-branches are derived from the parent population in parallel.

[0049] Large model voting decision unit: Based on the statistical summary of each branch, decide which branch to continue exploring or to perform backtracking.

[0050] Stagnation detection and penalty adjustment unit: When the search gets stuck in a local optimum, the complexity penalty is dynamically increased.

[0051] Large Model Interaction Interface: Responsible for communicating with the LLM API, constructing prompt words, parsing results, and handling exceptions.

[0052] Final Training and Evaluation Module: Evaluate the performance of the final model on the test set using the optimal subset of features.

[0053] This invention provides a feature selection method for high-dimensional medical image data based on population tree search and a hybrid intelligent genetic algorithm. The specific implementation is as follows: First, in the data preprocessing stage, the medical image feature dataset is loaded and divided into training and testing sets. Then, population tree search is initialized, constructing a search tree with a randomly generated initial subset of medical image features as the root node. Next, in the evaluation and fitness calculation stage, a fitness function is defined to evaluate the quality of each feature subset. This function comprehensively considers the model's predictive performance and a penalty term related to the number of selected features, aiming to balance accuracy and simplicity. Finally, in the branch evolution stage, multiple offspring branches are generated in parallel from the current node (parent population). Each branch is a new population generated through a hybrid genetic algorithm, where the genetic operations are primarily executed and guided by a large language model. Then, through intelligent decision-making and tree pruning, a large language model decision-maker analyzes the statistical information of all sub-branches and selects the most promising branch for the next round of exploration. If all branches perform poorly, a backtracking operation is performed, returning to the previous level node for re-exploration. Finally, the steps of branch evolution and intelligent decision-making are repeated, continuously performing a depth-first search on the population tree until a preset termination condition is met. The algorithm ultimately outputs the subset of features with the highest fitness found throughout the search process as the optimal solution.

[0054] Specifically, the present invention will be described in detail below with reference to the accompanying drawings. The specific execution flow of this method is as follows:

[0055] a. Parameters and Initialization: First, set the population size, number of branches, maximum number of tree levels, and penalty coefficient ( The data is then processed by the data loading and preprocessing module, and the population tree searcher creates the root node to initialize the first generation of the parent population.

[0056] b. Branching and Short-Range Evolution: Based on the current parent population, multiple branches are generated in parallel by the branching evolution unit. Within each branch, the hybrid genetic algorithm engine performs short-range evolution for a specified number of generations. This process prioritizes the use of the large model crossover / mutation unit, and the fitness evaluation unit calculates the fitness of each individual. After evolution is complete, elite individuals are retained and the historical solution cache is updated.

[0057] c. Stagnation Detection and Adjustment: The stagnation detection unit continuously monitors changes in the global optimum. If no improvement is achieved after several consecutive rounds, it enters "stagnation mode," increasing the complexity penalty coefficient. This encourages algorithms to escape local optima.

[0058] d. Large Model Voting and Tree Advancement / Backtracking: After all branches have completed short-range evolution, their statistical summaries (such as best / average fitness, number of features, etc.) are submitted to the large model voting decision unit. The LLM will analyze this information and return a decision instruction: if a valid branch number is returned, the algorithm explores one level down along that branch; if -1 is returned, a backtracking operation is performed, returning to the previous level node.

[0059] e. Termination and Output: Repeat steps b to d until the preset maximum number of tree levels is reached or other termination conditions are met. Finally, the final training and evaluation module trains and evaluates the model performance on an independent test set using a globally optimal subset of features, and outputs the final result.

[0060] To verify the effectiveness of this invention, the proposed method (GALLM) was compared with several traditional feature selection methods on multiple publicly available medical diagnostic datasets. As shown in Table 1, the proposed method achieved an AUC value of 0.9929 on the breast cancer diagnosis dataset, which is comparable to the best result (0.991 for the Sfs method). On the spect-heart and spectf-heart related datasets, the proposed method achieved AUC values ​​of 0.8097 and 0.8372, respectively, both significantly outperforming the other comparative methods.

[0061] Table 1. Performance comparison of the present invention with several traditional feature selection algorithms on baseline datasets.

[0062]

[0063] Experimental results show that the method of this invention can achieve optimal or near-optimal performance while selecting a smaller subset of features, resulting in higher prediction performance and lower feature cost. This is attributed to the balance between model performance and feature quantity in the fitness function, the dynamic penalty mechanism, and the synergistic effect of crossover / mutation and voting decisions guided by a large language model in multi-branch population tree search, which enables rapid convergence to a compact and efficient feature combination.

Claims

1. A medical image feature selection method based on population tree search and hybrid intelligent genetics, characterized in that: Includes the following steps: S1: Obtain the medical image feature dataset to be processed. The medical image feature dataset to be processed contains multiple samples, and each sample consists of a set of medical image quantitative features and corresponding diagnostic labels. S2: Initialize a population tree, where the root node contains multiple randomly generated initial medical image feature subsets, each feature subset serving as an individual in the population; S3: Evaluate and calculate fitness by evaluating the fitness of each subset of medical image features in the population through a preset fitness function. The fitness function takes into account the predictive performance of the diagnostic model trained on the subset of features and a penalty term related to the number of selected medical image features, in order to balance the accuracy and simplicity of the diagnostic model. S4: Perform branch evolution, starting from the current tree node, generating multiple child branches in parallel; each branch generates a new population by applying a hybrid genetic algorithm to the parent medical image feature subset population; the evolutionary operation of the hybrid genetic algorithm includes: a. Heuristic parent selection: Using a large language model (LLM), parent selection is performed based on the fitness index of candidate medical image feature subsets; b. Intelligent crossover: Prioritizes the use of the large language model to perform crossover operations on the parent generation medical image feature subset to generate a child generation medical image feature subset that combines the excellent characteristics of both parents; and reverts to the traditional crossover algorithm under preset conditions. c. Intelligent mutation: Prioritize the use of the large language model to perform mutation operations on the generated subset of medical image features to explore new feature combinations; and revert to the traditional mutation algorithm under preset conditions; S5: Perform intelligent decision-making and tree pruning. Utilize a decision model to analyze the statistical summary information of all child branches and select the most promising branch for the next round of evolution; or, if none of the branches have reached the preset potential threshold, perform a backtracking operation to return to the previous node. S6: Iteration and Termination. Repeat steps S4 and S5 to perform a depth-first search on the population tree until the preset termination condition is met. Finally, output the subset of medical image features with the highest fitness found during the entire search process as the optimal feature selection result.

2. The medical image feature selection method based on population tree search and hybrid intelligent genetics according to claim 1, characterized in that: The crossover and mutation operations of the Large Language Model (LLM) include: when processing high-dimensional feature sets, a block processing strategy is adopted: the complete feature vector is divided into multiple smaller blocks, each of which is processed independently by the Large Language Model (LLM), and finally the processing results of all blocks are recombined to form the final offspring individuals.

3. The medical image feature selection method based on population tree search and hybrid intelligent genetics according to claim 1, characterized in that: The large language model decision engine receives quantitative statistical data from each sub-branch population and outputs decision instructions for the next search direction based on preset prompts, thereby realizing intelligent pruning and navigation of the search tree.

4. The medical image feature selection method based on population tree search and hybrid intelligent genetics according to claim 1, characterized in that: The fitness function includes a dynamic penalty mechanism: when the search process fails to find a better solution for several consecutive generations, it is considered to be in a "stagnation mode", and the weight of the feature quantity penalty term is increased to encourage the algorithm to explore simpler feature combinations, thereby escaping local optima.

5. The medical image feature selection method based on population tree search and hybrid intelligent genetics according to claim 1, characterized in that: During the evolutionary process, the hybrid genetic algorithm maintains an elite pool containing historical best solutions. The historical information of the elite pool is provided as context to the large language model LLM, which helps it make more informative decisions when performing crossover and mutation operations.

Images